5D Library › One-Person Enterprise Automation, Avatar Livestream Funnels and Task-Exposure Arbitrage · 中文
Chapter 8. A Year’s Ledger and a Pre-Registered Test
One business, three ways to run it
Putting the parts together: a business with 5,500 task-hours a year (livestream operations, B2B delivery, product and content) and revenue from the avatar stream of Chapter 4, two retainers and one product. Setup A has a team and no AI agents; setup B is the founder alone with AI doing 0.8 of the task-hours; setup C adds a part-time editor who takes 400 hours of review and small tasks from the founder.
Calculation
# Chapter 8. p15: A year's ledger for three ways to run the same business.
# Work: 5,500 task-hours a year (livestream operations 1,500, B2B
# delivery 2,400, product and content 1,600). Revenue: avatar livestream
# RM 1,120 a day (p08), two retainers at RM 5,000 a month, a product at
# RM 6,000 a month. All three run the same avatar stream; they differ in
# who does the task-hours. A: founder plus a team, no AI agents (staff
# at RM 60,000 and 1,800 hours a year, 60 hours a year lost per pair of
# people). B: founder alone, AI doing 0.8 of the task-hours at RM 8 an
# hour, 0.05 review hours per AI hour. C: as B, plus a part-time editor
# (RM 24,000 a year) who takes 400 of the founder's hours.
H, HOURS, WAGE, K = 5500, 1800, 60000, 60
REVENUE = 1120 * 365 + 2 * 5000 * 12 + 6000 * 12
def team_for(need):
n = 1
while n * HOURS - K * n * (n - 1) / 2 < need:
n += 1
return n
if __name__ == "__main__":
print(" revenue RM {:,.0f} a year".format(REVENUE))
n = team_for(H)
a_cost = (n - 1) * WAGE
print(" A by hand : {} people incl. founder, cost RM {:,.0f}, profit"
" RM {:,.0f}".format(n, a_cost, REVENUE - a_cost))
s, r = 0.8, 0.05
founder = (1 - s) * H + r * s * H
b_cost = s * H * 8
print(" B solo+AI : founder {:,.0f} h, cost RM {:,.0f}, profit"
" RM {:,.0f}".format(founder, b_cost, REVENUE - b_cost))
c_cost = b_cost + 24000
print(" C +editor : founder {:,.0f} h, cost RM {:,.0f}, profit"
" RM {:,.0f}".format(founder - 400, c_cost, REVENUE - c_cost))
print(" founder hours spare in C: {:,.0f} of {:,}".format(
HOURS - (founder - 400), HOURS))
# revenue RM 600,800 a year
# A by hand : 4 people incl. founder, cost RM 180,000, profit RM 420,800
# B solo+AI : founder 1,320 h, cost RM 35,200, profit RM 565,600
# C +editor : founder 920 h, cost RM 59,200, profit RM 541,600
# founder hours spare in C: 880 of 1,800Revenue is RM 600,800 a year in all three. A needs 4 people including the founder, costs RM 180,000 and leaves RM 420,800. B costs RM 35,200 in AI fees and leaves RM 565,600, but the founder works 1,320 hours of the 1,800 available. C pays RM 24,000 more and leaves RM 541,600, with the founder at 920 hours and 880 hours spare. C is the more robust choice: B leaves too little slack for a launch, an illness or a queue of escalations (Chapter 2), and the spare hours in C are what a new product line or a third retainer needs.
A pre-registered test of the run-of-show
The livestream numbers of Chapters 3 and 4 are assumptions until measured. The test randomises hour-long blocks of a stream, not individual viewers, because viewers in the same hour share conditions; with about 200 viewers a block and an intra-block correlation of 0.005, the design effect is 1.995 (Kohavi et al. 2020).
Calculation
# Chapter 8. p16: How many viewers a fair test of the run-of-show needs.
# Two-proportion test, two-sided alpha 0.05, power 0.80. Tests: a 4- vs
# a 6-minute cycle (32.0 vs 22.9 orders per 1,000 viewers, p06); avatar
# vs human host (0.6 x 0.032 vs 0.032, p08); and a 10% relative lift on
# 0.032. Viewers arrive in hour-long blocks of about 200 that share
# conditions, so blocks, not viewers, are randomised; with an intra-block
# correlation of 0.005 the design effect is 1 + (200 - 1) x 0.005.
import math
from statistics import NormalDist
Z = NormalDist().inv_cdf
def n_per_arm(p1, p2, alpha=0.05, power=0.80):
pbar = (p1 + p2) / 2
num = (Z(1 - alpha / 2) * math.sqrt(2 * pbar * (1 - pbar)) +
Z(power) * math.sqrt(p1 * (1 - p1) + p2 * (1 - p2)))
return math.ceil(num ** 2 / (p1 - p2) ** 2)
if __name__ == "__main__":
deff = 1 + (200 - 1) * 0.005
print(" design effect {:.3f}".format(deff))
for name, p1, p2 in (("4 vs 6 minute cycle", 0.0320, 0.0229),
("avatar vs human", 0.6 * 0.032, 0.032),
("10% lift on 0.032", 0.032, 0.0352)):
n = n_per_arm(p1, p2)
blocks = math.ceil(n * deff / 200)
print(" {:20s}: {:7,d} viewers per arm, {:4d} hour-blocks".format(
name, n, blocks))
# design effect 1.995
# 4 vs 6 minute cycle : 5,060 viewers per arm, 51 hour-blocks
# avatar vs human : 2,389 viewers per arm, 24 hour-blocks
# 10% lift on 0.032 : 49,777 viewers per arm, 497 hour-blocksComparing a 4-minute with a 6-minute cycle (32.0 against 22.9 orders per 1,000 viewers in p06) needs 5,060 viewers per arm, about 51 hour-blocks. Comparing an avatar with a human host at a ratio of 0.6 needs 2,389 per arm, 24 blocks. A 10% relative lift on a conversion rate of 0.032 needs 49,777 viewers per arm, 497 blocks: small improvements cannot be confirmed by one small shop in a short time, and should not be claimed.
Note
A four-step check for a one-person enterprise. First, measure review hours per AI deliverable and the escalation share every week; they set the ceiling (p02, p04). Second, put a check after each agent step that faces a customer (p03). Third, compute break-even ROAS with the shop’s own refund losses, not the short rule (p07). Fourth, quit a salary only after the side business has covered living costs for several months, and keep at least one recurring client (p11, p12).
Hypothesis · not established
Pre-registered prediction (not yet tested): in a randomised test over hour-blocks, a 4-minute cycle yields at least 15% more orders per 1,000 viewers than a 6-minute cycle, and an avatar host converts at between 0.5 and 0.8 of a human host. What would show it wrong: the shorter cycle does not yield significantly more orders, or the avatar’s conversion ratio falls outside that range.
Summary of the book
AI agents shrink the best team size as coordination channels fall away: from 21 people to 4 at an AI share of 0.80, and to one person only when review is light (p01, p02). Errors compound along agent chains, and the founder answering escalations is a queue that saturates (p03, p04). A looping run-of-show has a best cycle length, about 4.0 minutes with illustrative exit rates, and viewers who arrive mid-demonstration buy least (p05, p06). The short break-even ROAS rule can mislead in both directions, and an avatar beats a human at peak unless it converts below 0.551 of a human (p07, p08). Exposure is a property of tasks; the safest income lies in tasks that are hard to automate and that automated tasks feed (p09, p10). Quit a salary only when the market has answered, and keep a recurring client in the portfolio (p11, p12). Bai Gui’s rule needs a market that swings back, and the Kaogongji’s crafts show what written standards save (p13, p14).
References
Fan, Y. (2026). One-Person Enterprise Automation, Avatar Livestream Funnels and Task-Exposure Arbitrage: Coasean Transaction-Cost Collapse, Cyclic Script Retention and High-Margin Solopreneur Portfolios [一人公司自动化架构、数字人直播漏斗与任务级人工智能套利] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090235 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).